mirrors/ryujinx
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

Implement IMUL, PCNT and CONT shader instructions, fix FFMA32I and HFMA32I (#2972)

Browse source 
* Implement IMUL shader instruction

* Implement PCNT/CONT instruction and fix FFMA32I

* Add HFMA232I to the table

* Shader cache version bump

* No Rc on Ffma32i
This commit is contained in:
gdkchan 2022-01-10 12:08:00 -03:00 • committed by GitHub
parent 952c6e4d45
commit 7f6b3d234a
No known key found for this signature in database
GPG key ID: 4AEE18F83AFDEB23
9 changed files with 108 additions and 68 deletions
Download patch file Download diff file Expand all files Collapse all files

42
Ryujinx.Graphics.Shader/Decoders/Decoder.cs
View file

@ -95,7 +95,7 @@ namespace Ryujinx.Graphics.Shader.Decoders
if (currBlock.OpCodes.Count != 0)
{
// We should have blocks for all possible branch targets,
// including those from SSY/PBK instructions.
// including those from PBK/PCNT/SSY instructions.
foreach (PushOpInfo pushOp in currBlock.PushOpCodes)
{
GetBlock(pushOp.Op.GetAbsoluteAddress());
@ -243,7 +243,7 @@ namespace Ryujinx.Graphics.Shader.Decoders
{
SetUserAttributeUses(config, op.Name, opCode);
}
else if (op.Name == InstName.Ssy || op.Name == InstName.Pbk)
else if (op.Name == InstName.Pbk || op.Name == InstName.Pcnt || op.Name == InstName.Ssy)
{
block.AddPushOp(op);
}
@ -512,8 +512,9 @@ namespace Ryujinx.Graphics.Shader.Decoders
private enum MergeType
{
Brk = 0,
Sync = 1
Brk,
Cont,
Sync
}
private struct PathBlockState
@ -629,7 +630,7 @@ namespace Ryujinx.Graphics.Shader.Decoders
for (int index = pushOpIndex; index < pushOpsCount; index++)
{
InstOp currentPushOp = current.PushOpCodes[index].Op;
MergeType pushMergeType = currentPushOp.Name == InstName.Ssy ? MergeType.Sync : MergeType.Brk;
MergeType pushMergeType = GetMergeTypeFromPush(currentPushOp.Name);
branchStack.Push((currentPushOp.GetAbsoluteAddress(), pushMergeType));
}
}
@ -643,9 +644,9 @@ namespace Ryujinx.Graphics.Shader.Decoders
}
InstOp lastOp = current.GetLastOp();
if (lastOp.Name == InstName.Sync || lastOp.Name == InstName.Brk)
if (IsPopBranch(lastOp.Name))
{
MergeType popMergeType = lastOp.Name == InstName.Sync ? MergeType.Sync : MergeType.Brk;
MergeType popMergeType = GetMergeTypeFromPop(lastOp.Name);
bool found = true;
ulong targetAddress = 0UL;
@ -662,7 +663,7 @@ namespace Ryujinx.Graphics.Shader.Decoders
(targetAddress, mergeType) = branchStack.Pop();
// Push the target address (this will be used to push the address
// back into the SSY/PBK stack when we return from that block),
// back into the PBK/PCNT/SSY stack when we return from that block),
Push(new PathBlockState(targetAddress, mergeType));
}
while (mergeType != popMergeType);
@ -705,5 +706,30 @@ namespace Ryujinx.Graphics.Shader.Decoders
}
}
}
public static bool IsPopBranch(InstName name)
{
return name == InstName.Brk || name == InstName.Cont || name == InstName.Sync;
}
private static MergeType GetMergeTypeFromPush(InstName name)
{
return name switch
{
InstName.Pbk => MergeType.Brk,
InstName.Pcnt => MergeType.Cont,
_ => MergeType.Sync
};
}
private static MergeType GetMergeTypeFromPop(InstName name)
{
return name switch
{
InstName.Brk => MergeType.Brk,
InstName.Cont => MergeType.Cont,
_ => MergeType.Sync
};
}
}
}

2
Ryujinx.Graphics.Shader/Decoders/InstDecoders.cs
View file

@ -1960,7 +1960,6 @@ namespace Ryujinx.Graphics.Shader.Decoders
public int Dest => (int)((_opcode >> 0) & 0xFF);
public int SrcA => (int)((_opcode >> 8) & 0xFF);
public int Imm32 => (int)(_opcode >> 20);
public int SrcC => (int)((_opcode >> 39) & 0xFF);
public int Pred => (int)((_opcode >> 16) & 0x7);
public bool PredInv => (_opcode & 0x80000) != 0;
public bool NegC => (_opcode & 0x200000000000000) != 0;
@ -2460,7 +2459,6 @@ namespace Ryujinx.Graphics.Shader.Decoders
public int Dest => (int)((_opcode >> 0) & 0xFF);
public int SrcA => (int)((_opcode >> 8) & 0xFF);
public int Imm => (int)(_opcode >> 20);
public int SrcC => (int)((_opcode >> 39) & 0xFF);
public HalfSwizzle ASwizzle => (HalfSwizzle)((_opcode >> 47) & 0x3);
public int Pred => (int)((_opcode >> 16) & 0x7);
public bool PredInv => (_opcode & 0x80000) != 0;

5
Ryujinx.Graphics.Shader/Decoders/InstTable.cs
View file

@ -55,7 +55,7 @@ namespace Ryujinx.Graphics.Shader.Decoders
Add("1110111110000xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Cctll, InstEmit.Cctll, InstProps.Ra);
Add("1110101111110xx0000000000000xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Cctlt, InstEmit.Cctlt);
Add("1110101111101xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Cctlt, InstEmit.Cctlt, InstProps.Rc);
Add("111000110101xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Cont, InstEmit.Cont);
Add("111000110101xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Cont, InstEmit.Cont, InstProps.Bra);
Add("0101000010011xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Cset, InstEmit.Cset, InstProps.Rd | InstProps.Ps);
Add("0101000010100xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Csetp, InstEmit.Csetp, InstProps.Pd | InstProps.Pdn | InstProps.Ps);
Add("0101000011001xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Cs2r, InstEmit.Cs2r, InstProps.Rd);
@ -101,7 +101,7 @@ namespace Ryujinx.Graphics.Shader.Decoders
Add("0011001x1xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Ffma, InstEmit.FfmaI, InstProps.Rd | InstProps.Ra | InstProps.Ib | InstProps.Rc);
Add("010010011xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Ffma, InstEmit.FfmaC, InstProps.Rd | InstProps.Ra | InstProps.Rc);
Add("010100011xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Ffma, InstEmit.FfmaRc, InstProps.Rd | InstProps.Ra | InstProps.Rc);
Add("000011xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Ffma32i, InstEmit.Ffma32i, InstProps.Rd | InstProps.Ra | InstProps.Rc);
Add("000011xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Ffma32i, InstEmit.Ffma32i, InstProps.Rd | InstProps.Ra);
Add("0101110000110xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Flo, InstEmit.FloR, InstProps.Rd | InstProps.Rb);
Add("0011100x00110xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Flo, InstEmit.FloI, InstProps.Rd | InstProps.Ib);
Add("0100110000110xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Flo, InstEmit.FloC, InstProps.Rd);
@ -129,6 +129,7 @@ namespace Ryujinx.Graphics.Shader.Decoders
Add("01110xxx0xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Hfma2, InstEmit.Hfma2I, InstProps.Rd | InstProps.Ra | InstProps.Ib | InstProps.Rc);
Add("01110xxx1xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Hfma2, InstEmit.Hfma2C, InstProps.Rd | InstProps.Ra | InstProps.Rc);
Add("01100xxx1xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Hfma2, InstEmit.Hfma2Rc, InstProps.Rd | InstProps.Ra | InstProps.Rc);
Add("0010100xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Hfma2, InstEmit.Hfma232i, InstProps.Rd | InstProps.Ra);
Add("0101110100001xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Hmul2, InstEmit.Hmul2R, InstProps.Rd | InstProps.Ra | InstProps.Rb);
Add("0111100x0xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Hmul2, InstEmit.Hmul2I, InstProps.Rd | InstProps.Ra | InstProps.Ib);
Add("0111100x1xxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxxx", InstName.Hmul2, InstEmit.Hmul2C, InstProps.Rd | InstProps.Ra);